The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes multiple Python tools and shell commands to manage the development lifecycle. This includes querying plugins via tools/plugin.py, performing benchmarks with tools/compute_utilization.py, managing metadata with tools/memory_manager.py, and using git for version control of the generated artifacts.
[DYNAMIC_EXECUTION]: The skill dynamically generates kernel.py and test_kernel.py by interpreting user PyTorch logic and learning from framework APIs. These generated scripts are subsequently executed to validate correctness and measure performance. Although execution occurs within a tools/sandbox.py, the runtime generation of executable code is a significant behavior.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it processes external, untrusted data to guide its primary logic.
Ingestion points: Reads user-supplied PyTorch logic, kernel_demo snippets, and plugin instructions located at .atrex_plugins/instructions.md.
Boundary markers: The instructions lack explicit delimitation or warnings to ignore embedded directives within the user-provided logic or plugin instructions.
Capability inventory: The agent can write to the local filesystem (kernel.py, test_kernel.py, memory/v0.json), query external knowledge tools via plugins, and execute shell commands to run tests and benchmarks.
Sanitization: There is no mention of sanitizing user input or plugin content before it is used to generate executable Python code.

gpu-kernel-baseline